AugMax Explained!

[@iskrabesamrtna]

I was wondering what kind of augmentation would you recommand for videos mapped with keypoints estimation on a person that exibits various gestures in multiclass classification (huge amount of classes) if the augmentation method is even used on such task. Any hint would do for my further research. Thanks!

[@connor-shorten]

All of the same augmentations in image classification should work for this since they preserve the label of the gestures. Please see RandAugment for a full list of these augmentations - if any of these might flip the label for your application you would want to remove it from the N which RandAugment samples from. You still may not want to do this since the added noise in the training may help and removing the augs could be a bit of a pain - compared to using the imgaug implementation off the shelf. Thank you for the question! Please let me know if this was clear or not useful, happy to help with further questions!

[@mrdbourke]

Fantastic video!

[@connor-shorten]

Thanks, I really appreciate it! Happy to help if you have any issues with Data Augmentation or Robustness Testing!

[@sayakpaul3152]

I think it was JS-Divergence for the consistency loss term. Very similar to KL-D but wanted to point out the detail.

[@connor-shorten]

Got you -- JS adds symmetry to KL, just looked it up haha

[@SakvaUA]

So what's preventing the augmentation weights from collapsing to 0 and just giving the original image on the output?

[@sayakpaul3152]

That is probably encouraged by the friendly adversarial training term in the loss. I am not 100% sure, though.

[@connor-shorten]

I have been experimenting with this -- the consistency loss is generally held up by the 1-hot class label loss. Collapse happens all the time if you try to drill this deeper into the representation or increase the weight of the loss.

[@connor-shorten]

I agree, entropy regularization also appears in MEMO: Test Time Robustness via Adaptation and Augmentation -- Marvin Zhang, Sergey Levine, and Chelsea Finn

[@dawwdd]

Good job as always

[@connor-shorten]

Thank you so much, I really appreciate it!

[@domenickmifsud]

Thanks! Great content

[@connor-shorten]

Thanks for watching!

[@rorrochanel]

Wow! Amazing!

[@connor-shorten]

Very exciting stuff, congrats to the authors! Thanks for watching!

[@MrMIB983]

Love long videos

[@connor-shorten]

Really glad to hear it, these are fun to make!